Add MIRAGE dataset: redundancy + temporal configs

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,89 @@
 ---
-license: mit
+license: other
+license_name: mixed-see-below
+license_link: https://github.com/DeepBio-Scientific/MIRAGE#provenance--licensing
+task_categories:
+  - tabular-regression
+tags:
+  - binding-affinity
+  - drug-discovery
+  - protein-ligand
+  - benchmark
+  - data-leakage
+pretty_name: MIRAGE
+configs:
+  - config_name: redundancy
+    data_files: redundancy/test-*.parquet
+    default: true
+  - config_name: temporal
+    data_files: temporal/test-*.parquet
 ---
+
+# MIRAGE: Measuring Interpolation and Redundancy in Affinity GEneralization
+
+A benchmark for measuring whether a protein–ligand **binding-affinity** model generalises
+beyond protein families that are heavily represented in the PDB, or whether its reported
+accuracy is carried by that redundancy.
+
+Frontier co-folding affinity models (Boltz-2, Nesso-1, …) report Pearson ≈ 0.6 on standard
+benchmarks. When the same models are stratified by protein-family size, that accuracy is
+concentrated on well-represented families and **collapses to ≈ 0 on singleton families** —
+exactly the case that matters when starting a program against a novel target.
+
+## Configs
+
+### `redundancy` (18,759 rows, default)
+PDBbind-derived protein–ligand complexes with experimental affinity, annotated with
+protein-family size so accuracy can be stratified by redundancy.
+
+| column | description |
+|---|---|
+| `id` | PDB code |
+| `sequence` | protein chain (single-letter) |
+| `smiles` | ligand SMILES (from the deposited structure) |
+| `pK` | experimental −log10(Kd/Ki/IC50); **higher = stronger** |
+| `affinity_type` | `Kd` / `Ki` / `IC50` |
+| `year` | PDB deposition year |
+| `family_id` | MMseqs2 30%-identity cluster representative |
+| `family_size` | number of complexes in that family (**the redundancy axis**) |
+| `redundancy_bin` | `1`, `2-5`, `6-20`, `21-80`, `81-300`, `301+` |
+| `ligand_nn_tanimoto` | ECFP4 nearest-neighbour similarity to any other ligand |
+| `in_core` | member of the balanced 3,360-target quick-eval subset |
+
+### `temporal` (649 rows)
+One structurally novel, post-2021-09-30 target (**OpenBind EV-A71 2A protease**), with
+measured KD and shipped baselines. Tests within-target ranking on data that no model with
+a ≤ 2021 cutoff could have seen.
+
+| column | description |
+|---|---|
+| `id` | Fragalysis code |
+| `sequence` | target protein |
+| `smiles` | ligand SMILES |
+| `pKD` | experimental −log10(KD); higher = stronger |
+| `baseline_molecular_weight`, `baseline_clogp`, `ref_boltz2` | reference predictions |
+
+## Usage
+
+```python
+import mirage
+
+def my_model(sequence, smiles):
+    return score               # higher = stronger binder
+
+print(mirage.run_redundancy(my_model).summary())
+print(mirage.run_temporal(my_model).summary())
+```
+
+See <https://github.com/DeepBio-Scientific/MIRAGE> for the harness, baselines, and CLI.
+
+## Provenance & licensing
+
+- **Redundancy set** is derived from **PDBbind** (protein sequences from the PDB; SMILES
+  from deposited ligands; affinities from PDBbind's curated literature values). Only derived
+  annotations are redistributed here, under MIT. Users should observe PDBbind's own terms
+  when using the underlying structures.
+- **Temporal set** is derived from the **OpenBind A71EV2A** release (CC0) and its public
+  benchmark repository.
+
+Family sizes are computed with MMseqs2 at 30% sequence identity / 80% coverage.

redundancy/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c50e8368ae75a2215501233bb4b2b0393757d6584f321577907e32d1506d968b
+size 4652560

temporal/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f525ff732c774a44b759488b926ef5d5b67b51a1857dfe9b1a20b78034f1fbbe
+size 32706